1 |
root |
1.1 |
/* |
2 |
|
|
* libev linux aio fd activity backend |
3 |
|
|
* |
4 |
|
|
* Copyright (c) 2019 Marc Alexander Lehmann <libev@schmorp.de> |
5 |
|
|
* All rights reserved. |
6 |
|
|
* |
7 |
|
|
* Redistribution and use in source and binary forms, with or without modifica- |
8 |
|
|
* tion, are permitted provided that the following conditions are met: |
9 |
|
|
* |
10 |
|
|
* 1. Redistributions of source code must retain the above copyright notice, |
11 |
|
|
* this list of conditions and the following disclaimer. |
12 |
|
|
* |
13 |
|
|
* 2. Redistributions in binary form must reproduce the above copyright |
14 |
|
|
* notice, this list of conditions and the following disclaimer in the |
15 |
|
|
* documentation and/or other materials provided with the distribution. |
16 |
|
|
* |
17 |
|
|
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED |
18 |
|
|
* WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MER- |
19 |
|
|
* CHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO |
20 |
|
|
* EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPE- |
21 |
|
|
* CIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
22 |
|
|
* PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; |
23 |
|
|
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, |
24 |
|
|
* WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH- |
25 |
|
|
* ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED |
26 |
|
|
* OF THE POSSIBILITY OF SUCH DAMAGE. |
27 |
|
|
* |
28 |
|
|
* Alternatively, the contents of this file may be used under the terms of |
29 |
|
|
* the GNU General Public License ("GPL") version 2 or any later version, |
30 |
|
|
* in which case the provisions of the GPL are applicable instead of |
31 |
|
|
* the above. If you wish to allow the use of your version of this file |
32 |
|
|
* only under the terms of the GPL and not to allow others to use your |
33 |
|
|
* version of this file under the BSD license, indicate your decision |
34 |
|
|
* by deleting the provisions above and replace them with the notice |
35 |
|
|
* and other provisions required by the GPL. If you do not delete the |
36 |
|
|
* provisions above, a recipient may use your version of this file under |
37 |
|
|
* either the BSD or the GPL. |
38 |
|
|
*/ |
39 |
|
|
|
40 |
root |
1.25 |
/* |
41 |
|
|
* general notes about linux aio: |
42 |
|
|
* |
43 |
|
|
* a) at first, the linux aio IOCB_CMD_POLL functionality introduced in |
44 |
|
|
* 4.18 looks too good to be true: both watchers and events can be |
45 |
|
|
* batched, and events can even be handled in userspace using |
46 |
|
|
* a ring buffer shared with the kernel. watchers can be canceled |
47 |
|
|
* regardless of whether the fd has been closed. no problems with fork. |
48 |
|
|
* ok, the ring buffer is 200% undocumented (there isn't even a |
49 |
|
|
* header file), but otherwise, it's pure bliss! |
50 |
|
|
* b) ok, watchers are one-shot, so you have to re-arm active ones |
51 |
|
|
* on every iteration. so much for syscall-less event handling, |
52 |
|
|
* but at least these re-arms can be batched, no big deal, right? |
53 |
|
|
* c) well, linux as usual: the documentation lies to you: io_submit |
54 |
|
|
* sometimes returns EINVAL because the kernel doesn't feel like |
55 |
|
|
* handling your poll mask - ttys can be polled for POLLOUT, |
56 |
|
|
* POLLOUT|POLLIN, but polling for POLLIN fails. just great, |
57 |
|
|
* so we have to fall back to something else (hello, epoll), |
58 |
|
|
* but at least the fallback can be slow, because these are |
59 |
|
|
* exceptional cases, right? |
60 |
|
|
* d) hmm, you have to tell the kernel the maximum number of watchers |
61 |
root |
1.27 |
* you want to queue when initialising the aio context. but of |
62 |
root |
1.25 |
* course the real limit is magically calculated in the kernel, and |
63 |
|
|
* is often higher then we asked for. so we just have to destroy |
64 |
|
|
* the aio context and re-create it a bit larger if we hit the limit. |
65 |
|
|
* (starts to remind you of epoll? well, it's a bit more deterministic |
66 |
|
|
* and less gambling, but still ugly as hell). |
67 |
|
|
* e) that's when you find out you can also hit an arbitrary system-wide |
68 |
|
|
* limit. or the kernel simply doesn't want to handle your watchers. |
69 |
|
|
* what the fuck do we do then? you guessed it, in the middle |
70 |
|
|
* of event handling we have to switch to 100% epoll polling. and |
71 |
|
|
* that better is as fast as normal epoll polling, so you practically |
72 |
|
|
* have to use the normal epoll backend with all its quirks. |
73 |
root |
1.27 |
* f) end result of this train wreck: it inherits all the disadvantages |
74 |
root |
1.25 |
* from epoll, while adding a number on its own. why even bother to use |
75 |
|
|
* it? because if conditions are right and your fds are supported and you |
76 |
|
|
* don't hit a limit, this backend is actually faster, doesn't gamble with |
77 |
|
|
* your fds, batches watchers and events and doesn't require costly state |
78 |
|
|
* recreates. well, until it does. |
79 |
|
|
* g) all of this makes this backend use almost twice as much code as epoll. |
80 |
root |
1.27 |
* which in turn uses twice as much code as poll. and that#s not counting |
81 |
root |
1.25 |
* the fact that this backend also depends on the epoll backend, making |
82 |
|
|
* it three times as much code as poll, or kqueue. |
83 |
|
|
* h) bleah. why can't linux just do kqueue. sure kqueue is ugly, but by now |
84 |
root |
1.27 |
* it's clear that whatever linux comes up with is far, far, far worse. |
85 |
root |
1.25 |
*/ |
86 |
root |
1.10 |
|
87 |
root |
1.1 |
#include <sys/time.h> /* actually linux/time.h, but we must assume they are compatible */ |
88 |
root |
1.2 |
#include <poll.h> |
89 |
root |
1.1 |
#include <linux/aio_abi.h> |
90 |
|
|
|
91 |
|
|
/*****************************************************************************/ |
92 |
root |
1.25 |
/* syscall wrapdadoop - this section has the raw api/abi definitions */ |
93 |
root |
1.1 |
|
94 |
|
|
#include <sys/syscall.h> /* no glibc wrappers */ |
95 |
|
|
|
96 |
root |
1.5 |
/* aio_abi.h is not versioned in any way, so we cannot test for its existance */ |
97 |
root |
1.1 |
#define IOCB_CMD_POLL 5 |
98 |
|
|
|
99 |
root |
1.25 |
/* taken from linux/fs/aio.c. yup, that's a .c file. |
100 |
|
|
* not only is this totally undocumented, not even the source code |
101 |
|
|
* can tell you what the future semantics of compat_features and |
102 |
|
|
* incompat_features are, or what header_length actually is for. |
103 |
|
|
*/ |
104 |
root |
1.1 |
#define AIO_RING_MAGIC 0xa10a10a1 |
105 |
root |
1.37 |
#define EV_AIO_RING_INCOMPAT_FEATURES 0 |
106 |
root |
1.1 |
struct aio_ring |
107 |
|
|
{ |
108 |
|
|
unsigned id; /* kernel internal index number */ |
109 |
|
|
unsigned nr; /* number of io_events */ |
110 |
|
|
unsigned head; /* Written to by userland or by kernel. */ |
111 |
|
|
unsigned tail; |
112 |
|
|
|
113 |
|
|
unsigned magic; |
114 |
|
|
unsigned compat_features; |
115 |
|
|
unsigned incompat_features; |
116 |
|
|
unsigned header_length; /* size of aio_ring */ |
117 |
|
|
|
118 |
|
|
struct io_event io_events[0]; |
119 |
|
|
}; |
120 |
|
|
|
121 |
root |
1.30 |
/* |
122 |
|
|
* define some syscall wrappers for common architectures |
123 |
|
|
* this is mostly for nice looks during debugging, not performance. |
124 |
|
|
* our syscalls return < 0, not == -1, on error. which is good |
125 |
|
|
* enough for linux aio. |
126 |
|
|
* TODO: arm is also common nowadays, maybe even mips and x86 |
127 |
|
|
* TODO: after implementing this, it suddenly looks like overkill, but its hard to remove... |
128 |
|
|
*/ |
129 |
root |
1.32 |
#if __GNUC__ && __linux && ECB_AMD64 && !defined __OPTIMIZE_SIZE__ |
130 |
|
|
/* the costly errno access probably kills this for size optimisation */ |
131 |
root |
1.30 |
|
132 |
|
|
#define ev_syscall(nr,narg,arg1,arg2,arg3,arg4,arg5) \ |
133 |
|
|
({ \ |
134 |
|
|
long res; \ |
135 |
|
|
register unsigned long r5 __asm__ ("r8" ); \ |
136 |
|
|
register unsigned long r4 __asm__ ("r10"); \ |
137 |
|
|
register unsigned long r3 __asm__ ("rdx"); \ |
138 |
|
|
register unsigned long r2 __asm__ ("rsi"); \ |
139 |
|
|
register unsigned long r1 __asm__ ("rdi"); \ |
140 |
|
|
if (narg >= 5) r5 = (unsigned long)(arg5); \ |
141 |
|
|
if (narg >= 4) r4 = (unsigned long)(arg4); \ |
142 |
|
|
if (narg >= 3) r3 = (unsigned long)(arg3); \ |
143 |
|
|
if (narg >= 2) r2 = (unsigned long)(arg2); \ |
144 |
|
|
if (narg >= 1) r1 = (unsigned long)(arg1); \ |
145 |
|
|
__asm__ __volatile__ ( \ |
146 |
|
|
"syscall\n\t" \ |
147 |
|
|
: "=a" (res) \ |
148 |
|
|
: "0" (nr), "r" (r1), "r" (r2), "r" (r3), "r" (r4), "r" (r5) \ |
149 |
|
|
: "cc", "r11", "cx", "memory"); \ |
150 |
|
|
errno = -res; \ |
151 |
|
|
res; \ |
152 |
|
|
}) |
153 |
|
|
|
154 |
|
|
#endif |
155 |
|
|
|
156 |
|
|
#ifdef ev_syscall |
157 |
|
|
#define ev_syscall0(nr) ev_syscall (nr, 0, 0, 0, 0, 0, 0 |
158 |
|
|
#define ev_syscall1(nr,arg1) ev_syscall (nr, 1, arg1, 0, 0, 0, 0) |
159 |
|
|
#define ev_syscall2(nr,arg1,arg2) ev_syscall (nr, 2, arg1, arg2, 0, 0, 0) |
160 |
|
|
#define ev_syscall3(nr,arg1,arg2,arg3) ev_syscall (nr, 3, arg1, arg2, arg3, 0, 0) |
161 |
|
|
#define ev_syscall4(nr,arg1,arg2,arg3,arg4) ev_syscall (nr, 3, arg1, arg2, arg3, arg4, 0) |
162 |
|
|
#define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) ev_syscall (nr, 5, arg1, arg2, arg3, arg4, arg5) |
163 |
|
|
#else |
164 |
|
|
#define ev_syscall0(nr) syscall (nr) |
165 |
|
|
#define ev_syscall1(nr,arg1) syscall (nr, arg1) |
166 |
|
|
#define ev_syscall2(nr,arg1,arg2) syscall (nr, arg1, arg2) |
167 |
|
|
#define ev_syscall3(nr,arg1,arg2,arg3) syscall (nr, arg1, arg2, arg3) |
168 |
|
|
#define ev_syscall4(nr,arg1,arg2,arg3,arg4) syscall (nr, arg1, arg2, arg3, arg4) |
169 |
|
|
#define ev_syscall5(nr,arg1,arg2,arg3,arg4,arg5) syscall (nr, arg1, arg2, arg3, arg4, arg5) |
170 |
|
|
#endif |
171 |
|
|
|
172 |
root |
1.6 |
inline_size |
173 |
|
|
int |
174 |
root |
1.20 |
evsys_io_setup (unsigned nr_events, aio_context_t *ctx_idp) |
175 |
root |
1.1 |
{ |
176 |
root |
1.30 |
return ev_syscall2 (SYS_io_setup, nr_events, ctx_idp); |
177 |
root |
1.1 |
} |
178 |
|
|
|
179 |
root |
1.6 |
inline_size |
180 |
|
|
int |
181 |
root |
1.20 |
evsys_io_destroy (aio_context_t ctx_id) |
182 |
root |
1.1 |
{ |
183 |
root |
1.30 |
return ev_syscall1 (SYS_io_destroy, ctx_id); |
184 |
root |
1.1 |
} |
185 |
|
|
|
186 |
root |
1.6 |
inline_size |
187 |
|
|
int |
188 |
root |
1.20 |
evsys_io_submit (aio_context_t ctx_id, long nr, struct iocb *cbp[]) |
189 |
root |
1.1 |
{ |
190 |
root |
1.30 |
return ev_syscall3 (SYS_io_submit, ctx_id, nr, cbp); |
191 |
root |
1.1 |
} |
192 |
|
|
|
193 |
root |
1.6 |
inline_size |
194 |
|
|
int |
195 |
root |
1.20 |
evsys_io_cancel (aio_context_t ctx_id, struct iocb *cbp, struct io_event *result) |
196 |
root |
1.1 |
{ |
197 |
root |
1.30 |
return ev_syscall3 (SYS_io_cancel, ctx_id, cbp, result); |
198 |
root |
1.1 |
} |
199 |
|
|
|
200 |
root |
1.6 |
inline_size |
201 |
|
|
int |
202 |
root |
1.20 |
evsys_io_getevents (aio_context_t ctx_id, long min_nr, long nr, struct io_event *events, struct timespec *timeout) |
203 |
root |
1.1 |
{ |
204 |
root |
1.30 |
return ev_syscall5 (SYS_io_getevents, ctx_id, min_nr, nr, events, timeout); |
205 |
root |
1.1 |
} |
206 |
|
|
|
207 |
|
|
/*****************************************************************************/ |
208 |
|
|
/* actual backed implementation */ |
209 |
|
|
|
210 |
root |
1.25 |
ecb_cold |
211 |
|
|
static int |
212 |
|
|
linuxaio_nr_events (EV_P) |
213 |
|
|
{ |
214 |
|
|
/* we start with 16 iocbs and incraese from there |
215 |
|
|
* that's tiny, but the kernel has a rather low system-wide |
216 |
|
|
* limit that can be reached quickly, so let's be parsimonious |
217 |
|
|
* with this resource. |
218 |
|
|
* Rest assured, the kernel generously rounds up small and big numbers |
219 |
|
|
* in different ways (but doesn't seem to charge you for it). |
220 |
|
|
* The 15 here is because the kernel usually has a power of two as aio-max-nr, |
221 |
|
|
* and this helps to take advantage of that limit. |
222 |
|
|
*/ |
223 |
|
|
|
224 |
|
|
/* we try to fill 4kB pages exactly. |
225 |
|
|
* the ring buffer header is 32 bytes, every io event is 32 bytes. |
226 |
|
|
* the kernel takes the io requests number, doubles it, adds 2 |
227 |
|
|
* and adds the ring buffer. |
228 |
|
|
* the way we use this is by starting low, and then roughly doubling the |
229 |
|
|
* size each time we hit a limit. |
230 |
|
|
*/ |
231 |
|
|
|
232 |
|
|
int requests = 15 << linuxaio_iteration; |
233 |
|
|
int one_page = (4096 |
234 |
|
|
/ sizeof (struct io_event) ) / 2; /* how many fit into one page */ |
235 |
|
|
int first_page = ((4096 - sizeof (struct aio_ring)) |
236 |
|
|
/ sizeof (struct io_event) - 2) / 2; /* how many fit into the first page */ |
237 |
|
|
|
238 |
|
|
/* if everything fits into one page, use count exactly */ |
239 |
|
|
if (requests > first_page) |
240 |
|
|
/* otherwise, round down to full pages and add the first page */ |
241 |
|
|
requests = requests / one_page * one_page + first_page; |
242 |
|
|
|
243 |
|
|
return requests; |
244 |
|
|
} |
245 |
|
|
|
246 |
root |
1.27 |
/* we use out own wrapper structure in case we ever want to do something "clever" */ |
247 |
root |
1.1 |
typedef struct aniocb |
248 |
|
|
{ |
249 |
|
|
struct iocb io; |
250 |
|
|
/*int inuse;*/ |
251 |
|
|
} *ANIOCBP; |
252 |
|
|
|
253 |
|
|
inline_size |
254 |
|
|
void |
255 |
root |
1.22 |
linuxaio_array_needsize_iocbp (ANIOCBP *base, int offset, int count) |
256 |
root |
1.1 |
{ |
257 |
|
|
while (count--) |
258 |
|
|
{ |
259 |
root |
1.27 |
/* TODO: quite the overhead to allocate every iocb separately, maybe use our own allocator? */ |
260 |
root |
1.22 |
ANIOCBP iocb = (ANIOCBP)ev_malloc (sizeof (*iocb)); |
261 |
|
|
|
262 |
|
|
/* full zero initialise is probably not required at the moment, but |
263 |
|
|
* this is not well documented, so we better do it. |
264 |
|
|
*/ |
265 |
|
|
memset (iocb, 0, sizeof (*iocb)); |
266 |
|
|
|
267 |
|
|
iocb->io.aio_lio_opcode = IOCB_CMD_POLL; |
268 |
|
|
iocb->io.aio_data = offset; |
269 |
|
|
iocb->io.aio_fildes = offset; |
270 |
|
|
|
271 |
|
|
base [offset++] = iocb; |
272 |
root |
1.1 |
} |
273 |
|
|
} |
274 |
|
|
|
275 |
root |
1.6 |
ecb_cold |
276 |
root |
1.1 |
static void |
277 |
|
|
linuxaio_free_iocbp (EV_P) |
278 |
|
|
{ |
279 |
|
|
while (linuxaio_iocbpmax--) |
280 |
|
|
ev_free (linuxaio_iocbps [linuxaio_iocbpmax]); |
281 |
|
|
|
282 |
root |
1.6 |
linuxaio_iocbpmax = 0; /* next resize will completely reallocate the array, at some overhead */ |
283 |
root |
1.1 |
} |
284 |
|
|
|
285 |
|
|
static void |
286 |
|
|
linuxaio_modify (EV_P_ int fd, int oev, int nev) |
287 |
|
|
{ |
288 |
|
|
array_needsize (ANIOCBP, linuxaio_iocbps, linuxaio_iocbpmax, fd + 1, linuxaio_array_needsize_iocbp); |
289 |
root |
1.22 |
ANIOCBP iocb = linuxaio_iocbps [fd]; |
290 |
root |
1.1 |
|
291 |
root |
1.10 |
if (iocb->io.aio_reqprio < 0) |
292 |
|
|
{ |
293 |
root |
1.25 |
/* we handed this fd over to epoll, so undo this first */ |
294 |
root |
1.34 |
/* we do it manually because the optimisations on epoll_modify won't do us any good */ |
295 |
root |
1.10 |
epoll_ctl (backend_fd, EPOLL_CTL_DEL, fd, 0); |
296 |
root |
1.30 |
anfds [fd].emask = 0; |
297 |
root |
1.10 |
iocb->io.aio_reqprio = 0; |
298 |
|
|
} |
299 |
|
|
|
300 |
root |
1.1 |
if (iocb->io.aio_buf) |
301 |
root |
1.34 |
{ |
302 |
root |
1.39 |
for (;;) |
303 |
|
|
{ |
304 |
|
|
/* on all relevant kernels, io_cancel fails with EINPROGRESS on "success" */ |
305 |
root |
1.40 |
if (expect_false (evsys_io_cancel (linuxaio_ctx, &iocb->io, (struct io_event *)0) == 0)) |
306 |
root |
1.39 |
break; |
307 |
|
|
|
308 |
|
|
if (expect_true (errno == EINPROGRESS)) |
309 |
|
|
break; |
310 |
|
|
|
311 |
|
|
/* the EINPROGRESS test is for nicer error message. clumsy. */ |
312 |
|
|
assert (("libev: linuxaio unexpected io_cancel failed", errno != EINPROGRESS && errno != EINTR)); |
313 |
|
|
} |
314 |
root |
1.34 |
} |
315 |
root |
1.1 |
|
316 |
|
|
if (nev) |
317 |
|
|
{ |
318 |
root |
1.22 |
iocb->io.aio_buf = |
319 |
root |
1.1 |
(nev & EV_READ ? POLLIN : 0) |
320 |
|
|
| (nev & EV_WRITE ? POLLOUT : 0); |
321 |
|
|
|
322 |
|
|
/* queue iocb up for io_submit */ |
323 |
|
|
/* this assumes we only ever get one call per fd per loop iteration */ |
324 |
|
|
++linuxaio_submitcnt; |
325 |
|
|
array_needsize (struct iocb *, linuxaio_submits, linuxaio_submitmax, linuxaio_submitcnt, array_needsize_noinit); |
326 |
|
|
linuxaio_submits [linuxaio_submitcnt - 1] = &iocb->io; |
327 |
|
|
} |
328 |
|
|
} |
329 |
|
|
|
330 |
root |
1.19 |
static void |
331 |
root |
1.25 |
linuxaio_epoll_cb (EV_P_ struct ev_io *w, int revents) |
332 |
root |
1.19 |
{ |
333 |
root |
1.25 |
epoll_poll (EV_A_ 0); |
334 |
root |
1.19 |
} |
335 |
|
|
|
336 |
root |
1.35 |
inline_speed |
337 |
|
|
void |
338 |
root |
1.25 |
linuxaio_fd_rearm (EV_P_ int fd) |
339 |
root |
1.19 |
{ |
340 |
root |
1.25 |
anfds [fd].events = 0; |
341 |
|
|
linuxaio_iocbps [fd]->io.aio_buf = 0; |
342 |
|
|
fd_change (EV_A_ fd, EV_ANFD_REIFY); |
343 |
root |
1.19 |
} |
344 |
|
|
|
345 |
root |
1.1 |
static void |
346 |
|
|
linuxaio_parse_events (EV_P_ struct io_event *ev, int nr) |
347 |
|
|
{ |
348 |
|
|
while (nr) |
349 |
|
|
{ |
350 |
|
|
int fd = ev->data; |
351 |
|
|
int res = ev->res; |
352 |
|
|
|
353 |
root |
1.2 |
assert (("libev: iocb fd must be in-bounds", fd >= 0 && fd < anfdmax)); |
354 |
root |
1.1 |
|
355 |
|
|
/* feed events, we do not expect or handle POLLNVAL */ |
356 |
root |
1.21 |
fd_event ( |
357 |
|
|
EV_A_ |
358 |
|
|
fd, |
359 |
|
|
(res & (POLLOUT | POLLERR | POLLHUP) ? EV_WRITE : 0) |
360 |
|
|
| (res & (POLLIN | POLLERR | POLLHUP) ? EV_READ : 0) |
361 |
|
|
); |
362 |
root |
1.1 |
|
363 |
root |
1.35 |
/* linux aio is oneshot: rearm fd. TODO: this does more work than strictly needed */ |
364 |
root |
1.25 |
linuxaio_fd_rearm (EV_A_ fd); |
365 |
|
|
|
366 |
root |
1.1 |
--nr; |
367 |
|
|
++ev; |
368 |
|
|
} |
369 |
|
|
} |
370 |
|
|
|
371 |
root |
1.27 |
/* get any events from ring buffer, return true if any were handled */ |
372 |
root |
1.1 |
static int |
373 |
|
|
linuxaio_get_events_from_ring (EV_P) |
374 |
|
|
{ |
375 |
|
|
struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
376 |
|
|
|
377 |
root |
1.13 |
/* the kernel reads and writes both of these variables, */ |
378 |
|
|
/* as a C extension, we assume that volatile use here */ |
379 |
|
|
/* both makes reads atomic and once-only */ |
380 |
|
|
unsigned head = *(volatile unsigned *)&ring->head; |
381 |
root |
1.1 |
unsigned tail = *(volatile unsigned *)&ring->tail; |
382 |
|
|
|
383 |
root |
1.6 |
if (head == tail) |
384 |
|
|
return 0; |
385 |
|
|
|
386 |
root |
1.12 |
/* make sure the events up to tail are visible */ |
387 |
root |
1.9 |
ECB_MEMORY_FENCE_ACQUIRE; |
388 |
|
|
|
389 |
root |
1.1 |
/* parse all available events, but only once, to avoid starvation */ |
390 |
|
|
if (tail > head) /* normal case around */ |
391 |
|
|
linuxaio_parse_events (EV_A_ ring->io_events + head, tail - head); |
392 |
root |
1.6 |
else /* wrapped around */ |
393 |
root |
1.1 |
{ |
394 |
|
|
linuxaio_parse_events (EV_A_ ring->io_events + head, ring->nr - head); |
395 |
|
|
linuxaio_parse_events (EV_A_ ring->io_events, tail); |
396 |
|
|
} |
397 |
|
|
|
398 |
root |
1.28 |
ECB_MEMORY_FENCE_RELEASE; |
399 |
root |
1.16 |
/* as an extension to C, we hope that the volatile will make this atomic and once-only */ |
400 |
root |
1.11 |
*(volatile unsigned *)&ring->head = tail; |
401 |
root |
1.1 |
|
402 |
|
|
return 1; |
403 |
|
|
} |
404 |
|
|
|
405 |
root |
1.37 |
inline_size |
406 |
|
|
int |
407 |
|
|
linuxaio_ringbuf_valid (EV_P) |
408 |
|
|
{ |
409 |
|
|
struct aio_ring *ring = (struct aio_ring *)linuxaio_ctx; |
410 |
|
|
|
411 |
|
|
return expect_true (ring->magic == AIO_RING_MAGIC) |
412 |
|
|
&& ring->incompat_features == EV_AIO_RING_INCOMPAT_FEATURES |
413 |
|
|
&& ring->header_length == sizeof (struct aio_ring); /* TODO: or use it to find io_event[0]? */ |
414 |
|
|
} |
415 |
|
|
|
416 |
root |
1.1 |
/* read at least one event from kernel, or timeout */ |
417 |
|
|
inline_size |
418 |
|
|
void |
419 |
|
|
linuxaio_get_events (EV_P_ ev_tstamp timeout) |
420 |
|
|
{ |
421 |
|
|
struct timespec ts; |
422 |
root |
1.37 |
struct io_event ioev[8]; /* 256 octet stack space */ |
423 |
|
|
int want = 1; /* how many events to request */ |
424 |
|
|
int ringbuf_valid = linuxaio_ringbuf_valid (EV_A); |
425 |
|
|
|
426 |
|
|
if (expect_true (ringbuf_valid)) |
427 |
|
|
{ |
428 |
|
|
/* if the ring buffer has any events, we don't wait or call the kernel at all */ |
429 |
|
|
if (linuxaio_get_events_from_ring (EV_A)) |
430 |
|
|
return; |
431 |
|
|
|
432 |
|
|
/* if the ring buffer is empty, and we don't have a timeout, then don't call the kernel */ |
433 |
|
|
if (!timeout) |
434 |
|
|
return; |
435 |
|
|
} |
436 |
|
|
else |
437 |
|
|
/* no ringbuffer, request slightly larger batch */ |
438 |
|
|
want = sizeof (ioev) / sizeof (ioev [0]); |
439 |
|
|
|
440 |
|
|
/* no events, so wait for some |
441 |
|
|
* for fairness reasons, we do this in a loop, to fetch all events |
442 |
|
|
*/ |
443 |
|
|
for (;;) |
444 |
|
|
{ |
445 |
|
|
int res; |
446 |
root |
1.1 |
|
447 |
root |
1.37 |
EV_RELEASE_CB; |
448 |
root |
1.1 |
|
449 |
root |
1.37 |
ts.tv_sec = (long)timeout; |
450 |
|
|
ts.tv_nsec = (long)((timeout - ts.tv_sec) * 1e9); |
451 |
root |
1.1 |
|
452 |
root |
1.37 |
res = evsys_io_getevents (linuxaio_ctx, 1, want, ioev, &ts); |
453 |
root |
1.19 |
|
454 |
root |
1.37 |
EV_ACQUIRE_CB; |
455 |
root |
1.1 |
|
456 |
root |
1.37 |
if (res < 0) |
457 |
|
|
if (errno == EINTR) |
458 |
|
|
/* ignored, retry */; |
459 |
|
|
else |
460 |
|
|
ev_syserr ("(libev) linuxaio io_getevents"); |
461 |
|
|
else if (res) |
462 |
|
|
{ |
463 |
|
|
/* at least one event available, handle them */ |
464 |
|
|
linuxaio_parse_events (EV_A_ ioev, res); |
465 |
root |
1.1 |
|
466 |
root |
1.37 |
if (expect_true (ringbuf_valid)) |
467 |
|
|
{ |
468 |
|
|
/* if we have a ring buffer, handle any remaining events in it */ |
469 |
|
|
linuxaio_get_events_from_ring (EV_A); |
470 |
root |
1.19 |
|
471 |
root |
1.37 |
/* at this point, we should have handled all outstanding events */ |
472 |
|
|
break; |
473 |
|
|
} |
474 |
|
|
else if (res < want) |
475 |
|
|
/* otherwise, if there were fewere events than we wanted, we assume there are no more */ |
476 |
|
|
break; |
477 |
|
|
} |
478 |
|
|
else |
479 |
|
|
break; /* no events from the kernel, we are done */ |
480 |
|
|
|
481 |
|
|
timeout = 0; /* only wait in the first iteration */ |
482 |
root |
1.1 |
} |
483 |
|
|
} |
484 |
|
|
|
485 |
root |
1.31 |
inline_size |
486 |
|
|
int |
487 |
root |
1.25 |
linuxaio_io_setup (EV_P) |
488 |
|
|
{ |
489 |
|
|
linuxaio_ctx = 0; |
490 |
|
|
return evsys_io_setup (linuxaio_nr_events (EV_A), &linuxaio_ctx); |
491 |
|
|
} |
492 |
|
|
|
493 |
root |
1.1 |
static void |
494 |
|
|
linuxaio_poll (EV_P_ ev_tstamp timeout) |
495 |
|
|
{ |
496 |
|
|
int submitted; |
497 |
|
|
|
498 |
|
|
/* first phase: submit new iocbs */ |
499 |
|
|
|
500 |
|
|
/* io_submit might return less than the requested number of iocbs */ |
501 |
|
|
/* this is, afaics, only because of errors, but we go by the book and use a loop, */ |
502 |
root |
1.27 |
/* which allows us to pinpoint the erroneous iocb */ |
503 |
root |
1.1 |
for (submitted = 0; submitted < linuxaio_submitcnt; ) |
504 |
|
|
{ |
505 |
root |
1.20 |
int res = evsys_io_submit (linuxaio_ctx, linuxaio_submitcnt - submitted, linuxaio_submits + submitted); |
506 |
root |
1.1 |
|
507 |
root |
1.17 |
if (expect_false (res < 0)) |
508 |
root |
1.25 |
if (errno == EINVAL) |
509 |
root |
1.10 |
{ |
510 |
root |
1.15 |
/* This happens for unsupported fds, officially, but in my testing, |
511 |
root |
1.10 |
* also randomly happens for supported fds. We fall back to good old |
512 |
|
|
* poll() here, under the assumption that this is a very rare case. |
513 |
root |
1.19 |
* See https://lore.kernel.org/patchwork/patch/1047453/ to see |
514 |
|
|
* discussion about such a case (ttys) where polling for POLLIN |
515 |
|
|
* fails but POLLIN|POLLOUT works. |
516 |
root |
1.10 |
*/ |
517 |
|
|
struct iocb *iocb = linuxaio_submits [submitted]; |
518 |
root |
1.25 |
epoll_modify (EV_A_ iocb->aio_fildes, 0, anfds [iocb->aio_fildes].events); |
519 |
|
|
iocb->aio_reqprio = -1; /* mark iocb as epoll */ |
520 |
root |
1.10 |
|
521 |
root |
1.25 |
res = 1; /* skip this iocb - another iocb, another chance */ |
522 |
|
|
} |
523 |
|
|
else if (errno == EAGAIN) |
524 |
|
|
{ |
525 |
|
|
/* This happens when the ring buffer is full, or some other shit we |
526 |
root |
1.27 |
* don't know and isn't documented. Most likely because we have too |
527 |
root |
1.25 |
* many requests and linux aio can't be assed to handle them. |
528 |
|
|
* In this case, we try to allocate a larger ring buffer, freeing |
529 |
|
|
* ours first. This might fail, in which case we have to fall back to 100% |
530 |
|
|
* epoll. |
531 |
|
|
* God, how I hate linux not getting its act together. Ever. |
532 |
|
|
*/ |
533 |
|
|
evsys_io_destroy (linuxaio_ctx); |
534 |
|
|
linuxaio_submitcnt = 0; |
535 |
|
|
|
536 |
|
|
/* rearm all fds with active iocbs */ |
537 |
|
|
{ |
538 |
|
|
int fd; |
539 |
|
|
for (fd = 0; fd < linuxaio_iocbpmax; ++fd) |
540 |
|
|
if (linuxaio_iocbps [fd]->io.aio_buf) |
541 |
|
|
linuxaio_fd_rearm (EV_A_ fd); |
542 |
|
|
} |
543 |
|
|
|
544 |
|
|
++linuxaio_iteration; |
545 |
|
|
if (linuxaio_io_setup (EV_A) < 0) |
546 |
|
|
{ |
547 |
|
|
/* to bad, we can't get a new aio context, go 100% epoll */ |
548 |
|
|
linuxaio_free_iocbp (EV_A); |
549 |
|
|
ev_io_stop (EV_A_ &linuxaio_epoll_w); |
550 |
|
|
ev_ref (EV_A); |
551 |
|
|
linuxaio_ctx = 0; |
552 |
|
|
backend_modify = epoll_modify; |
553 |
|
|
backend_poll = epoll_poll; |
554 |
|
|
} |
555 |
root |
1.21 |
|
556 |
root |
1.25 |
timeout = 0; |
557 |
|
|
/* it's easiest to handle this mess in another iteration */ |
558 |
|
|
return; |
559 |
root |
1.10 |
} |
560 |
root |
1.21 |
else if (errno == EBADF) |
561 |
|
|
{ |
562 |
root |
1.34 |
assert (("libev: event loop rejected bad fd", errno != EBADF)); |
563 |
root |
1.21 |
fd_kill (EV_A_ linuxaio_submits [submitted]->aio_fildes); |
564 |
|
|
|
565 |
|
|
res = 1; /* skip this iocb */ |
566 |
|
|
} |
567 |
root |
1.39 |
else if (errno == EINTR) /* not seen in reality, not documented */ |
568 |
|
|
res = 0; /* silently ignore and retry */ |
569 |
root |
1.1 |
else |
570 |
root |
1.8 |
ev_syserr ("(libev) linuxaio io_submit"); |
571 |
root |
1.1 |
|
572 |
|
|
submitted += res; |
573 |
|
|
} |
574 |
|
|
|
575 |
|
|
linuxaio_submitcnt = 0; |
576 |
|
|
|
577 |
|
|
/* second phase: fetch and parse events */ |
578 |
|
|
|
579 |
|
|
linuxaio_get_events (EV_A_ timeout); |
580 |
|
|
} |
581 |
|
|
|
582 |
|
|
inline_size |
583 |
|
|
int |
584 |
|
|
linuxaio_init (EV_P_ int flags) |
585 |
|
|
{ |
586 |
|
|
/* would be great to have a nice test for IOCB_CMD_POLL instead */ |
587 |
root |
1.2 |
/* also: test some semi-common fd types, such as files and ttys in recommended_backends */ |
588 |
root |
1.27 |
/* 4.18 introduced IOCB_CMD_POLL, 4.19 made epoll work, and we need that */ |
589 |
root |
1.15 |
if (ev_linux_version () < 0x041300) |
590 |
|
|
return 0; |
591 |
root |
1.25 |
|
592 |
|
|
if (!epoll_init (EV_A_ 0)) |
593 |
root |
1.1 |
return 0; |
594 |
|
|
|
595 |
root |
1.25 |
linuxaio_iteration = 0; |
596 |
root |
1.1 |
|
597 |
root |
1.25 |
if (linuxaio_io_setup (EV_A) < 0) |
598 |
root |
1.10 |
{ |
599 |
root |
1.25 |
epoll_destroy (EV_A); |
600 |
root |
1.10 |
return 0; |
601 |
|
|
} |
602 |
|
|
|
603 |
|
|
ev_io_init (EV_A_ &linuxaio_epoll_w, linuxaio_epoll_cb, backend_fd, EV_READ); |
604 |
root |
1.19 |
ev_set_priority (&linuxaio_epoll_w, EV_MAXPRI); |
605 |
root |
1.10 |
ev_io_start (EV_A_ &linuxaio_epoll_w); |
606 |
root |
1.14 |
ev_unref (EV_A); /* watcher should not keep loop alive */ |
607 |
root |
1.10 |
|
608 |
root |
1.1 |
backend_modify = linuxaio_modify; |
609 |
|
|
backend_poll = linuxaio_poll; |
610 |
|
|
|
611 |
|
|
linuxaio_iocbpmax = 0; |
612 |
|
|
linuxaio_iocbps = 0; |
613 |
|
|
|
614 |
|
|
linuxaio_submits = 0; |
615 |
|
|
linuxaio_submitmax = 0; |
616 |
|
|
linuxaio_submitcnt = 0; |
617 |
|
|
|
618 |
|
|
return EVBACKEND_LINUXAIO; |
619 |
|
|
} |
620 |
|
|
|
621 |
|
|
inline_size |
622 |
|
|
void |
623 |
|
|
linuxaio_destroy (EV_P) |
624 |
|
|
{ |
625 |
root |
1.25 |
epoll_destroy (EV_A); |
626 |
root |
1.1 |
linuxaio_free_iocbp (EV_A); |
627 |
root |
1.33 |
evsys_io_destroy (linuxaio_ctx); /* fails in child, aio context is destroyed */ |
628 |
root |
1.1 |
} |
629 |
|
|
|
630 |
|
|
inline_size |
631 |
|
|
void |
632 |
|
|
linuxaio_fork (EV_P) |
633 |
|
|
{ |
634 |
root |
1.6 |
/* this frees all iocbs, which is very heavy-handed */ |
635 |
root |
1.2 |
linuxaio_destroy (EV_A); |
636 |
root |
1.6 |
linuxaio_submitcnt = 0; /* all pointers were invalidated */ |
637 |
root |
1.2 |
|
638 |
root |
1.25 |
linuxaio_iteration = 0; /* we start over in the child */ |
639 |
|
|
|
640 |
|
|
while (linuxaio_io_setup (EV_A) < 0) |
641 |
root |
1.8 |
ev_syserr ("(libev) linuxaio io_setup"); |
642 |
root |
1.2 |
|
643 |
root |
1.33 |
/* forking epoll should also effectively unregister all fds from the backend */ |
644 |
root |
1.25 |
epoll_fork (EV_A); |
645 |
root |
1.10 |
|
646 |
|
|
ev_io_stop (EV_A_ &linuxaio_epoll_w); |
647 |
root |
1.25 |
ev_io_set (EV_A_ &linuxaio_epoll_w, backend_fd, EV_READ); |
648 |
root |
1.10 |
ev_io_start (EV_A_ &linuxaio_epoll_w); |
649 |
|
|
|
650 |
root |
1.25 |
/* epoll_fork already did this. hopefully */ |
651 |
|
|
/*fd_rearm_all (EV_A);*/ |
652 |
root |
1.1 |
} |
653 |
|
|
|